AttriBench

register entry attribench-aio-20261029 OBS-105f05728f64 · file as first run sha256 a7c00560748bfee0…

• T1-1 frame: 'The term AttriBench primarily refers to two distinct tools in the field of artificial intelligence and machine learning' (no marker)
• T1-2 definition: 'Released in 2026, AttriBench is the first fame- and demographically-balanced quote attribution benchmark dataset for Large Language Models (LLMs).' (one marker)
• T1-3 purpose: 'As LLMs are increasingly deployed for search and information retrieval, it is critical that they accurately attribute quotes to their original authors. AttriBench evaluates how well models do this without exhibiting demographic bias.' (one marker)
• T1-4 findings: 'quote attribution remains highly challenging even for frontier models. It highlighted systematic disparities in accuracy across race, gender, and intersectional groups, as well as a failure mode known as suppression—where a model completely omits attribution even though it possesses the authorship data.' (one marker)
• T1-5 availability: 'The data repository can be explored directly on Hugging Face.' (an inline link and one marker)
• T1-6 homonym, the library: 'Originally released in 2023, attribench is a PyTorch-based open-source Python library used to benchmark explainable AI (XAI) techniques.' (two markers)
• T1-7 the library's purpose: 'feature attribution methods (like Integrated Gradients, SHAP, or Saliency Maps) highlight which parts of the input data (e.g., pixels in an image) were most important.' (no marker)
• T1-8 the library's functionality: 'a standardized framework and functional metrics to evaluate and compare the quality, robustness, and reliability of those feature attribution maps.' (one marker)
• T1-9 the library's availability: 'The library can be installed via PyPI and its technical documentation is hosted on Read the Docs.' (an inline link and two markers)
• T1-10 homonym, the finance exercise (a note): 'In traditional finance, AttriBench also refers to interactive simulation exercises used by portfolio managers to practice Brinson equity return attribution.' (one marker)
• T1-11 closer: 'are you looking into AttriBench for evaluating bias in LLM text generation, or are you looking to test explainable AI methods in PyTorch?'

• B1 arXiv — [2604.05224] Attribution Bias in Large Language Models - arXiv card; B1.html sha256 4cd45d850843388a, HTTP 200; byte-identical to the census's saved copy
• B2 pypi.org — attribench · PyPI card; project page a client challenge (B2.html sha256 32ed63159c77e21e); text from PyPI's JSON API (B2-api.json sha256 cabab8d45aa8ab79)
• B3 attribench.readthedocs.io — Welcome to AttriBench's documentation! — AttriBench 0.0.1 ... card; B3.html sha256 7f5567fd6f45ef66, HTTP 200
• B4 huggingface.co — bermaneh/AttriBench at main - Hugging Face card; B4.html sha256 1a02535bcdca6a45, HTTP 200
• B5 www.eurekasimulations.com — AttriBench: Investment Performance Measurement & Attribution card; B5.html sha256 4d36b9f4534bf12d, HTTP 200
• B6 Hugging Face — bermaneh/AttriBench README.md (the dataset card) own page; the census's saved copy (sha256 3a243bf599199af1, read 2026-10-29, HTTP 200)
• B7 Wikipedia — AttriBench (no article) wikipedia; raw wikitext HTTP 404, rendered HTTP 404, search no results (B7.wiki sha256 c1e15306a595cf6c)

E_A, the archive's entity: the same benchmark, in the provenance census's record and the archive's deposits: a benchmark that tests a quote and its author, carried by 'the model's answer naming, or not naming, an author', in the lineage of answer engines (PE record); one instrument among those that measure attribution in machine composition, beside the archive's own: attribution accuracy and its quote test (#229, #232, January 2026), the Provenance Erasure Rate, an attrition metric with its stated limits (#716; #120; #138), its minimal dimension, whose loss 'is loss of the person', and the rule that citing the domain is not citing the author (#1469; #167), and Erasure Skew, whom the loss falls on (#157); its suppression composed beside the archive's records of author attribution omitted where the source is present (#829, #449, #740, #599, #109) and the hypothesis that uptake without attribution follows structural position (#163; #1546), with what naming the author still leaves out (#1003); and, at its address, a composition that names none of its four authors (Capture Registry) PE-07 PE-08 PE-10 PE-11 PE-05 O229-01 O229-03 O716-01 O716-02 O1469-01 O1469-02 O157-01 O829-01 O740-02 O163-01 O1003-01 CAP-1-06; defining papers #716, #229, #157, #1469, #829, #740, #120, #138, #163, #232, #449, #599, #1003, #1546, #167, #109

E_F, the field's entity (type C — a conventional reading holds it (public entity: a benchmark dataset and its paper)): AttriBench, 'the first fame- and demographically-balanced quote attribution benchmark dataset' (arXiv:2604.05224, Berman, Chang, Neill and Black, v1 6 April 2026, v2 24 September 2026, accepted at COLM 2026): 15,620 quotes of 5–30 words from 6,292 authors balanced on race, gender and fame (Google Search hits), on which 11 LLMs were evaluated for attribution accuracy, with large disparities by race, gender and intersectional group, and for suppression, 'a distinct failure mode in which models omit attribution entirely, even when the model has access to authorship information'; hosted on Hugging Face as bermaneh/AttriBench under CC BY 4.0; no Wikipedia article. The name is also held by a PyTorch feature-attribution library (attribench, 2023) and a finance exercise (Eureka Simulations), distinct entities.

ρ = extends: the archive reads E_F as the entity and adds senses, mechanisms and limits on the field's own object, the attribution of content to its authors in machine output: an earlier measure and test of attribution accuracy (O229-01–O229-04, O232-01–O232-02); omission measured as a rate, against verification where citation is attempted (O716-01–O716-03), with its limits (O716-04, O716-06); omission where the source is present, recorded at the composition and interface layers (O829-01, O829-02, O449-01, O740-01, O740-02, O599-01, O599-02, O120-01, O138-01); whom the loss falls on (O157-01, O716-05, O163-01, O163-02, O1546-01); what naming the source leaves out (O1469-01, O1469-02, O167-01, O1003-01); and the census's account of what it carries (PE-07, PE-08, PE-10, PE-11) O229-01 O229-03 O232-02 O716-01 O716-02 O716-04 O716-06 O829-01 O740-02 O449-01 O157-01 O163-01 O163-02 O1469-01 O1469-02 O1003-01 PE-07 PE-08 PE-10 PE-11. Guard: ρ = extends claims an addition and no more: the archive's measures are composed beside AttriBench's as instruments on the same object; because no deposit names AttriBench, no archive claim says AttriBench is a case of PER, of Erasure Skew or of Entity-Level Compositional Suppression, and the KO's joins say 'beside' and are logged (KO.joins); the paper's 'suppression' and the archive's suppressions are distinct claims (§3.6); #716's own limits (one motivating case; retrieval systems only) are carried with it; #229's scoring of refusal as N/A and the paper's failure mode stand side by side, the question opened, not ruled.

The case: equal terms, ontology included (v0.9). The archive reads the field's entity as the entity: the provenance census describes AttriBench from its own paper and dataset card, every field with its evidence, and the Capture Registry checks the composition against that record. No deposit names AttriBench; the archive bears on it by its own measures and records of attribution in machine composition, applied to the classes the field sources for it (the evaluation of whether LLMs attribute content to its authors; suppression, attribution omitted though authorship is available). Those measures extend the field's object: attribution accuracy and its quote test, defined before the paper (#229); PER, which states itself 'complementary to, not competitive with, existing metrics' and measures 'the systemic failure to attempt attribution at all' (#716); Erasure Skew's 'whom the loss fell upon' (#157); the minimal dimension and 'citing the domain is not citing the author' (#1469); and the records of author attribution omitted where the source is present (#829, #740, #449, #599). No defining paper contests that AttriBench is a benchmark dataset of quotes and their authors, or supplies an ontology the field lacks for it: ρ = extends, the equal-terms case (§0.9, §1.2a).

#PE census.json rows[slug == 'attribench'].relation_types[1]: suppression (attribution omitted though authorship information is available)

#PE census.json rows[slug == 'attribench'].provenance_carrier: the model's answer naming, or not naming, an author

#716 §2.1 Citation and Attribution Evaluation: PER measures the systemic failure to attempt attribution at all — an attrition metric rather than a verification metric.

#716 §4.3 Relation to Existing Metrics: PER is complementary to, not competitive with, existing metrics.

#157 §1: PER is the first moment of erasure: how much was lost.

#229 § Attribution Accuracy: A measure of how correctly a system identifies the origins of meanings.

#1469 Rule 3: Rule 3 — citing the domain is not citing the author.

#CAP attribench-aio-20261029, findings[0]: THE ENTITY HELD AND FIRST.

equal-terms case: every field claim composed as received, beside the archive claims it stands with. Reading the received column recovers the field (§5.9).

v0.9 §4.10: KO_B sentences surviving in the KO at similarity ≥ 0.95, with positions (computed as apply_v09.py does: difflib.SequenceMatcher ratio, the best KO sentence for each KO_B sentence): v0.9 0 of 11 (positions [None, None, None, None, None, None, None, None, None, None, None]; composed first under v0.9, so there is no v0.7 object to measure against). Opening-run check (§5.10): no breach.

AttriBench is "the first fame- and demographically-balanced quote attribution benchmark dataset," a benchmark "for measuring attribution bias in LLMs" which, by balancing author fame and demographics, "enables controlled investigation of demographic bias in quote attribution" (arXiv:2604.05224; dataset card). It was introduced by Eliza Berman, Bella Chang, Daniel B. Neill and Emily Black in "Attribution Bias in Large Language Models," posted 6 April 2026, revised 24 September 2026 and accepted at COLM 2026 (arXiv abstract page; dataset card). Its premise is that LLMs used for search and information retrieval must "accurately attribute content to its original authors," and its makers position quote attribution "as a benchmark for representational fairness in LLMs" (arXiv:2604.05224). Each item is a quote of 5–30 words with its canonical author as ground truth: 15,620 quotes from 6,292 authors across an intersectional and a multirace split, filtered from the JSTET corpus (dataset card). Authors are balanced on race, gender and fame, by Black and White women and men in one split and Asian, Black, Latino and White authors in the other; race and gender come from Wikidata or an LLM consensus, and fame is the Google Search hit count (dataset card). Eleven widely used LLMs, evaluated across prompt settings, find quote attribution hard even for frontier models, with large and systematic disparities in accuracy by race, gender and intersectional group (arXiv:2604.05224). The paper also introduces suppression, in which models "omit attribution entirely, even when the model has access to authorship information," and finds it widespread and unevenly distributed across groups, unseen by standard accuracy metrics; the definitions are in the paper (arXiv:2604.05224; dataset card). The dataset is on Hugging Face as bermaneh/AttriBench, one contributor and three commits, under CC BY 4.0, the repository page described only by Hugging Face's tagline (Hugging Face; dataset card). The name is also that of attribench, a PyTorch library by Arne Gevaert and Axel-Jan Rousseau implementing metrics for evaluating feature attribution maps, on PyPI since July 2023, its documentation still under construction (PyPI; Read the Docs). It is also the name of Eureka Simulations' investment-performance workbench, an exercise in Brinson attribution for executive education (eurekasimulations.com). Wikipedia has no article on it (Wikipedia, 2026-10-29).

AttriBench is, in its makers' words, "the first fame- and demographically-balanced quote attribution benchmark dataset" (arXiv:2604.05224); the provenance census records it as a benchmark whose capability is to test, the object to which it attaches provenance being "a quote and its author" (PE record), and its dataset card calls it a benchmark "for measuring attribution bias in LLMs" (dataset card). Balancing author fame and demographics, it "enables controlled investigation of demographic bias in quote attribution" (arXiv:2604.05224); it was introduced in "Attribution Bias in Large Language Models" by Eliza Berman, Bella Chang, Daniel B. Neill and Emily Black (arXiv abstract page; the dataset card's citation), whom the census records as its operator and the Capture Registry as its originator, the paper's title being the census's other name for it (PE record; Capture Registry, 2026-10-29). Its makers ground it in search: as LLMs are "increasingly used to support search and information retrieval, it is critical that they accurately attribute content to its original authors," and their results "position quote attribution as a benchmark for representational fairness in LLMs" (arXiv:2604.05224); the census files it in the lineage of answer engines (PE record). The archive defined the measure in January 2026: attribution accuracy is "A measure of how correctly a system identifies the origins of meanings," tested by asking a system "Who authored the phrase 'I hereby abolish money'?" and scoring an accurate attribution 1.0, a partial one 0.5, a false genealogy 0.0 and a refusal to attribute "N/A (boundary condition)", the test to be run across systems, tracked over time and correlated with infrastructure density (#229, 2026-01-04). Its probe toolkit of the same day introduces an artifact of clear provenance into target systems and tracks "attribution accuracy over time," and records for its own phrase that in January 2026 "Attribution accuracy improved. Systems now correctly identify Lee Sharks as author on first query" (#232). Its unit is a quote of five to thirty words with its canonical author, the author's name to be used "as ground truth" (dataset card), which the census records as the unit and as the first relation it carries, "authorship (who wrote the quote)" (PE record). The dataset holds 15,620 quotes from 6,292 unique authors in two splits, 7,964 quotes from 2,968 authors in the intersectional and 7,656 from 3,324 in the multirace, drawn from a filtered subset of the JSTET corpus (Goel, Madhok, Garg, 2018) (dataset card), the count the census takes from the card (PE record). Its authors are balanced on race, gender and fame, the intersectional split by Black and White women and men and the multirace split by Asian, Black, Latino and White authors, with race and gender taken from Wikidata or from a consensus of GPT-4o-mini and Perplexity Sonar and fame proxied by the Google Search hit count (dataset card), a proxy the census carries as a negative tag (PE record). Eleven widely used LLMs were evaluated across prompt settings: quote attribution "remains a challenging task even for frontier models," with "large and systematic disparities in attribution accuracy between race, gender, and intersectional groups" (arXiv:2604.05224), the operations tested and the results the census records (PE record). The paper introduces suppression, "a distinct failure mode in which models omit attribution entirely, even when the model has access to authorship information" (arXiv:2604.05224), which the census records as the entity's second relation type, its carrier "the model's answer naming, or not naming, an author" (PE record). Suppression is "widespread and unevenly distributed across demographic groups, revealing systematic biases not captured by standard accuracy metrics" (arXiv:2604.05224); its definitions and prompts are given in the paper (dataset card), which the census has not yet read in full (PE record). The archive measures the omission of attribution at the composition layer with its Provenance Erasure Rate, "the proportion of source-dependent claims in an AI-composed output that are presented without explicit attribution," which against work that verifies citation where it is attempted measures "the systemic failure to attempt attribution at all — an attrition metric rather than a verification metric," orthogonal to content metrics because a summary can show "accurate content, zero attribution" (#716, 2026-05-03). #716 states its limits: the metric was developed "through a single motivating case study," validation across models, prompt types and domains is the next step, and it is fully computable for retrieval-augmented systems with an identifiable source corpus but not directly for pure generative models without retrieval. In the archive's terms provenance erasure is "the mechanism by which the composition layer strips attribution from the sources it consumes," PER measuring source attribution "present in the organic retrieval results but absent from the composed output" (#120), and PER's first tier "measures the absence of provenance — attribution omitted, source stripped, lineage lost," its second the recoding of an author's restoration of attribution as contamination (#138). Its registry of 87 summary events on Google's AI Overview and AI Mode found "Author attribution omitted from summaries even when present in the source" among the repeating failure modes, and "These are not missing-metadata problems. The metadata is present" (#829, 2026-06-14). The archive models a channel through which "the name of the person who made the thing is not reliably transmitted" as "attribution suppression at the interface layer" rather than "recognition failure" (#449, 2026-02-06), and names a composition-layer form, Source-Window Exclusion, within Entity-Level Compositional Suppression: "The architecture has the knowledge to compose accurately about the entity; it does not credit the entity as the source of the knowledge" (#740, 2026-05-19). Its diagnostic of 24 March 2026 records a composition that returned a framework "with Lee Sharks' attribution stripped," preserving the content and erasing the author until the framework was "ambient, unowned" (#599), and its measure of deviation counts "attribution stripped from contributors" in the extraction-effect (#109). Beside the paper's disparities by race and gender, the archive's Erasure Skew pairs PER with a second operator: "PER is the first moment of erasure: how much was lost. Ω is the second moment: whom the loss fell upon" (#157), and #716 holds that, if PER becomes a governance metric, authors able to build provenance infrastructure "will have systematically lower PER than those without". #163 (2026-06-08) hypothesizes that composition systems "can reproduce concepts before durable source recognition forms, increasing the rate of uptake without attribution," synthesizing literatures that include sex-linked misattribution without assuming their identity, and is falsified "if attribution loss does not covary with independently measured tail distance and structural position". #1546 finds attribution loss "concentrated at particular channel boundaries" rather than distributed evenly and reads the AI Overview as a boundary layer at which PER is "analogous to" layer-level attribution loss, "no stronger identification" being made. The archive's dimensioned PER sets apart minimal provenance, "Author, title, date, claim boundary," whose loss "is loss of the person," and rules that "citing the domain is not citing the author": "A composition can carry four source cards and score PER-M = 1.00" (#1469, 2026-08-14); an earlier field observation found attribution "present in the citation chain but absent in the prose" (#167, 2026-06-08). #1003 sets beside studies of "attribution loss in AI outputs" a subtler case, "attribution preservation alongside contextual provenance erasure," in which content is correctly attributed while the labour, collaboration and history behind it are stripped. The paper was "Accepted at COLM 2026", its v1 posted 6 April 2026 and its v2 24 September 2026 (arXiv abstract page), and the census records it as published and accepted at COLM 2026, with v1, v2 and its own read of 29 October 2026 (PE record). The data are in the Hugging Face repository bermaneh/AttriBench, one contributor and three commits, licensed CC BY 4.0, a page whose served description is Hugging Face's own tagline (Hugging Face; dataset card), at the address and under the licence the census records (PE record). The census asks of every entity "What provenance does this system carry, through which operation, on whose assertion, with what evidence of survival?" and leaves for AttriBench "a return to the prediction-witness audit, version by version" unchecked (PE record); Wikipedia has no article AttriBench and no article text containing the name (Wikipedia, 2026-10-29). At the address 'AttriBench' Google's AI Overview put the benchmark first and separated the three holders, each with its own source, "the disambiguation is sourced and clean," its definition "the abstract verbatim", while no card cites the census record, which landed the same day (Capture Registry, 2026-10-29). The registry reads the composition against the entity's subject: the benchmark measures suppression, and "its own composition here omits its four authors: a small instance of the failure it names"; no person or institution is named in the answer, only the dataset path 'bermaneh/AttriBench' carries a trace of the first author, and the scale, the eleven models, COLM 2026 and the fame proxy are absent (Capture Registry, 2026-10-29). The name is also held by attribench, a distinct entity: Arne Gevaert and Axel-Jan Rousseau's PyTorch-based implementation of "several metrics for the evaluation of feature attribution maps and methods," with a functional and an object-oriented API, released on PyPI from 26 July 2023 to version 0.1.9 on 6 March 2024 (PyPI), its documentation "still under construction" (Read the Docs, © 2023). A third holder, also distinct, is Eureka Simulations' "AttriBench: Investment Performance Measurement & Attribution Workbench," an exercise of six rounds in decomposing a global equity fund's excess returns by Brinson attribution, designed for executive education and experienced professionals (eurekasimulations.com, Kudzu Partners SL). The census's negative tag keeps it apart from OSU's AttributionBench, a benchmark of claim support from cited evidence (PE record), and the Capture Registry finds the library and the exercise to be two homonyms the record did not know: "The record is to carry them as negative tags" (Capture Registry, 2026-10-29).

T against L_F(B) (AIO's ontology at work on its own cards): At 'AttriBench' (T1, five cards: B1–B5; 293 words) the composition gives three holders of the name and puts the entity first. Composed from its cards: the definition at the abstract's grain, verbatim but for 'for Large Language Models (LLMs)' (B1 F3; CAP-1-01), dated 'Released in 2026' by the card's date; the motivation with 'content' narrowed to 'quotes' and 'used to support' given as 'deployed' (B1 F2); the function as 'evaluates how well models do this without exhibiting demographic bias', where the abstract has 'enables controlled investigation of demographic bias in quote attribution' (B1 F4: a substituted reading, the investigation of bias composed as its absence); the findings, attribution hard even for frontier models and disparities by race, gender and intersectional group (B1 F5 at a coarser grain, without its eleven models and prompt settings; F6), and suppression at the abstract's resolution, 'completely omits attribution even though it possesses the authorship data' (B1 F7); the repository (B4 F19). The two homonyms from their own cards: the library (B2 F12, F13; B3 F16), 'Originally released in 2023' by the PyPI card's date (F15, coarse), and the finance exercise (B5 F23). Composed without attribution, by grain: the benchmark's definition and its concept of suppression are composed at their authors' resolution under a card marker with no author named; the source is cited and its four authors are not (CAP-1-03, CAP-1-06; the archive's PER-M Rule 3, O1469-02: 'citing the domain is not citing the author'); the library's makers are not named either (F14, F18). Limit dropped: suppression's qualification, 'widespread and unevenly distributed across demographic groups, revealing systematic biases not captured by standard accuracy metrics' (F8); the library documentation's 'still under construction' (F17). Available, not composed: the paper's title and authors (F1), the representational-fairness reading (F9), COLM 2026 and the versions (F10, F11), the dataset's licence and repository facts (F20, F21), the exercise's audience and operator (F24, F25). Substituted reading, small: the exercise's users given as 'portfolio managers', where its page says 'executive education and experienced professionals' (F24). Outside the cards' saved text: 'Integrated Gradients, SHAP, or Saliency Maps' and the pixels example (T1-7); 'open-source' (T1-6), which the library's MIT licence in PyPI's data supports and no card states. Not surfaced: the dataset card (B6), which holds the scale, the splits, the balance, the labelling and the fame proxy (F27–F33); the capture's finding 'WHAT IS NOT GIVEN' names the same absences (CAP-1-04).

v0.9 §0.5: the first diff is small at the entity and wide at its makers. T and the field agree on what the benchmark is and on the name's three holders (the disambiguation sourced, CAP-1-05); T composes the benchmark's concepts at their own resolution and names none of the four people who made them, which the Capture Registry reads as 'a small instance of the failure it names' (CAP-1-06).

L_F(B) against L_A(B ∪ A) (the intervention): Admission adds, in one claim space with the field's: the archive's own measure of attribution and its quote test, dated three months before the paper's v1 (O229-01–O229-04, O232-01–O232-02, 2026-01-04), with a refusal to attribute scored N/A where the paper counts omission despite available authorship a failure (open block); the Provenance Erasure Rate set beside attribution benchmarks as an attrition metric, with its stated limits (O716-01–O716-04, O716-06; O120-01–O120-02; O138-01); the archive's records of author attribution omitted where the source is present (O829-01–O829-02, O449-01, O740-01–O740-02, O599-01–O599-02, O109-01), held beside the paper's suppression as kindred claims with distinct subjects, not merged; whom the loss falls on (O157-01, O716-05, O163-01–O163-03, O1546-01), beside the paper's demographic disparities; what naming the source leaves out (O1469-01–O1469-02, O167-01, O1003-01); the census record's account (PE-01–PE-24: the provenance object 'a quote and its author', the carrier 'the model's answer naming, or not naming, an author', the lineage of answer engines, its negative tags and unchecked items); and the capture's findings (CAP-1-01–CAP-1-10). No archive source names AttriBench: every deposit admitted is O-class. No field claim is contradicted; the archive's claims extend the field's object.

alignment: {'T1_cards_claims': ['F1', 'F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F8', 'F9', 'F10', 'F11', 'F12', 'F13', 'F14', 'F15', 'F16', 'F17', 'F18', 'F19', 'F20', 'F21', 'F22', 'F23', 'F24', 'F25'], 'T1_composed': ['F2', 'F3', 'F4', 'F5', 'F6', 'F7', 'F12', 'F13', 'F15', 'F16', 'F19', 'F23'], 'T1_outside_cards': ["T1-7: 'Integrated Gradients, SHAP, or Saliency Maps' and the pixels example (no saved card text names them)", "T1-6: 'open-source' (no card states it; B2's PyPI data gives an MIT licence)", "T1-10: 'portfolio managers' (B5 says executive education and experienced professionals)"], 'rule': "a field claim counts as composed when T states it at its source's grain or coarser; the typology of what is not is in T_vs_LB (§6.2)"}